GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

GraphForge. Training Working Agents with Graph-Anchored Workspace Synthesis. By Qisheng Su and eleven others, from the University of Science and Technology of China, Fudan University, and two other institutions. Published on the thirtieth of September. No code repository is linked to it.

A working agent has to read diverse files, coordinate tools, and produce deliverables. Training one calls for tasks built on many real files, with results that can be verified, and few pipelines synthesize that kind of data. The ones that exist fall short in opposite directions. Some generate the files with models, and the files lack realism and diversity. Others build tasks on real files but without verifiers specific to the task, so the quality of a result is left unchecked. Realism without verification, or verification of something unrealistic: neither is the training data a working agent needs.

GraphForge grounds both halves, the task and its verification, in real files. The device that does it is an evidence graph built over the files in a workspace. Because the task statement and the grading rubrics are both derived from that graph, every requirement is backed by files the agent can actually open, and every rubric criterion is anchored to the files needed to check it.

Generation starts from seeds grounded in occupations, which gives controlled diversity. For each seed, GraphForge assembles a workspace of real files and builds the evidence graph over their relations. The task and rubrics come out of the graph. An initial rollout then tests whether the task can actually be executed, and a revision agent repairs the task and rubrics against the original files. Only after that are trajectories collected. The order matters: a task is checked against reality twice, once when it is derived from the graph and again when a real attempt is made at it, before any agent is trained on it.

The authors fine-tune Qwen three point six, twenty-seven B, on two thousand one hundred and sixty-nine GraphForge trajectories. Under OpenHands, GDPVal rises to one thousand four hundred forty-five point seven, a gain of sixty-five point seven.

Under Claude Code, Workspace-Bench-Lite reaches sixty-three point seven, up seven point seven. SpreadsheetBench II reaches twenty-four point zero, up thirteen point seven. A further step, rejection fine-tuning on the fine-tuned model's own rollouts, uses the evidence-anchored rubrics to select which candidates to keep, and it improves all three benchmarks again. The authors read that as a sign the rubrics give a useful selection signal. The data and models are available.

One graph does two jobs in this pipeline. It decides what a task asks for, and it decides how an answer is judged, so the requirement and the check can never drift apart: both point at the same real files. That shared anchoring is why the rubrics earn a second role after the data is built. First they define what a correct trajectory looks like. Later, in rejection fine-tuning, they choose which of the model's own attempts deserve to be learned from.